$2,893 of compute, and where it went.

Every model call in the study, priced at the providers' published API rates: 2.33 billion tokens over 200 agent-hours, in 34 research sessions, 24 Claude Code sessions with their subagents, and 1,844 benchmark runs. Running the same research and benchmark yourself would cost about this much before counting your own time.

Where the $2,893 went

Claude priced per model at Anthropic's published rates, with cache reads and cache writes at their own rates; other models at their providers' rates. The research sessions' transcripts are not public: their cost is estimated from the length of each document, the search and fetch budgets the documents record, and the number of subagents, at Fable 5 prices.

Show the numbers

Part of the work	Cost at API prices
Research sessions behind the 34 research documents (estimated from the reports)	$1,483
Reviews, writing, building and analysis, in Claude Code sessions	$801
Benchmark runs of Claude Code	$542
Other models: graders, labellers and task writers	$53
Claude helper calls inside the benchmark	$13
Total	$2,893

When the money was spent

Benchmark runsClaude Code sessions

Measured spend only: this project's Claude Code sessions, dated by each reply, and the benchmark runs, dated by the day their campaign run started. The research sessions (about $1,483) have no dates and are left out, as are the other models. The busiest day was 15 Oct ($267).

Show the numbers

Day	Claude Code sessions	Benchmark runs	Total
16 Sep	$101.39	$0.00	$101.39
17 Sep	$0.00	$0.00	$0.00
18 Sep	$0.00	$0.00	$0.00
19 Sep	$0.00	$0.00	$0.00
20 Sep	$0.00	$0.00	$0.00
21 Sep	$0.00	$0.00	$0.00
22 Sep	$0.00	$0.00	$0.00
23 Sep	$0.00	$0.00	$0.00
24 Sep	$15.50	$0.00	$15.50
25 Sep	$22.41	$0.00	$22.41
26 Sep	$0.00	$0.00	$0.00
27 Sep	$0.00	$0.00	$0.00
28 Sep	$0.00	$0.00	$0.00
29 Sep	$0.00	$0.00	$0.00
30 Sep	$0.00	$0.00	$0.00
1 Oct	$14.32	$0.00	$14.32
2 Oct	$0.00	$0.00	$0.00
3 Oct	$0.00	$0.00	$0.00
4 Oct	$166.82	$0.00	$166.82
5 Oct	$121.17	$30.50	$151.67
6 Oct	$0.00	$0.00	$0.00
7 Oct	$0.00	$0.00	$0.00
8 Oct	$14.38	$35.70	$50.08
9 Oct	$109.53	$0.00	$109.53
10 Oct	$25.23	$57.48	$82.71
11 Oct	$49.03	$70.41	$119.44
12 Oct	$61.92	$137.98	$199.90
13 Oct	$0.00	$0.00	$0.00
14 Oct	$0.00	$0.00	$0.00
15 Oct	$56.92	$210.11	$267.03
16 Oct	$0.00	$0.00	$0.00
17 Oct	$0.00	$0.00	$0.00
18 Oct	$0.00	$0.00	$0.00
19 Oct	$42.86	$0.00	$42.86

Most tokens are cache reads; most of the bill is not

Cache readsCache writesOutput

Measured Claude usage only: this project's Claude Code sessions and every benchmark run, 2.08 billion tokens and $1,340 at API prices; the estimated research sessions have no breakdown and are left out. Cache reads are cheap per token, so they dominate the token count but not the bill; writing the cache and the model's own output are where most of the money goes. Uncached input was under 0.1% of both.

Show the numbers

Token type	Tokens	Share of tokens	Cost	Share of cost
Cache reads	1,996.3 m	95.8%	$417	31.1%
Cache writes	73.8 m	3.5%	$630	47.0%
Output	13.4 m	0.6%	$293	21.8%
Input	0.0 m	0.0%	$0	0.0%

Cheap tokens in bulk and dear tokens in small numbers cost about the same

Measured Claude usage only (this project's Claude Code sessions and every benchmark run), by token type: how many tokens, against what each million cost on average. Along a dashed line the cost is the same; the area of each circle is its cost. Cache reads outnumber output tokens 150 to 1 but cost 105 times less each, which puts the two on nearly the same cost line: $417 against $293. Cache writes, $630, cost the most.

Show the numbers

Token type	Tokens	Cost	Per million tokens
Cache reads	1,996.3 m	$417	$0.21
Cache writes	73.8 m	$630	$8.54
Output	13.4 m	$293	$21.92
Input	0.0 m	$0	$5.69

The first request of a session costs several times any later one

Cache writesCache readsOutputUncached input

The mean cost of each request in the repository screen's 312 sessions, by its place in the session, while at least 30 sessions were still running. The first request costs about $0.09, 97% of it for writing the whole prompt prefix (system prompt, tools, any repository text) to the cache at the one-hour rate. Every later request reads that prefix back at a fortieth of the price of writing it, and writes only what is new, so it costs about $0.024. Standing context is paid in full once per session, then cheaply on every request; short sessions pay the most for it per unit of work.

Show the numbers

Request	Cache writes	Cache reads	Output	Input	Total	Sessions
1	$0.0855	$0.0020	$0.0003	$0.0000	$0.0878	312
2	$0.0339	$0.0042	$0.0003	$0.0000	$0.0384	312
3	$0.0236	$0.0050	$0.0002	$0.0000	$0.0288	312
4	$0.0452	$0.0056	$0.0002	$0.0000	$0.0510	312
5	$0.0185	$0.0068	$0.0008	$0.0000	$0.0261	309
6	$0.0171	$0.0075	$0.0007	$0.0000	$0.0253	250
7	$0.0163	$0.0079	$0.0005	$0.0000	$0.0248	223
8	$0.0150	$0.0085	$0.0008	$0.0000	$0.0243	200
9	$0.0122	$0.0092	$0.0009	$0.0000	$0.0224	175
10	$0.0112	$0.0101	$0.0009	$0.0000	$0.0221	143
11	$0.0104	$0.0108	$0.0008	$0.0000	$0.0220	123
12	$0.0099	$0.0114	$0.0007	$0.0000	$0.0220	104
13	$0.0087	$0.0120	$0.0006	$0.0000	$0.0212	86
14	$0.0086	$0.0130	$0.0004	$0.0000	$0.0220	70
15	$0.0104	$0.0137	$0.0005	$0.0000	$0.0246	59
16	$0.0103	$0.0139	$0.0003	$0.0000	$0.0246	53
17	$0.0104	$0.0148	$0.0002	$0.0000	$0.0254	43
18	$0.0066	$0.0152	$0.0011	$0.0000	$0.0229	35

Research

Item	Count
Research sessions, one per research document	34
Words in the research documents	192,785
Web searches and page fetches by those sessions	about 1,600 and 900
Sources cited, deduplicated	1,040
Public repositories reviewed from source, each with a written verdict	46 (36 in full, 10 in part)

Benchmark

Item	Count
Configurations specified for testing	107
Benchmark runs of Claude Code	1,844
Tokens used by those runs	411,692,001
Agent-hours of those runs	49.4
Tasks in the final harm-check bank	69

Most tokens are cached context that the model re-reads on every turn; they are cheap per token but dominate the count. The research sessions searched and read the web; the Claude Code sessions reviewed the repositories, wrote the setup and the benchmark, and analysed the results; the benchmark runs are Claude Code building the test tasks.